Researchers from Iowa State University and ETH Zürich have developed a mathematical framework called “rulebooks” that enables autonomous robots to navigate conflicting safety directives by ranking priorities rather than blending them into a single score.
The system, detailed in a recent study published by IEEE Transactions on Robotics, addresses a fundamental flaw in how self-driving cars and drones currently calculate risk during unavoidable emergencies.
As autonomous robots operate with decreasing human oversight, they increasingly encounter complex environments where adhering to one safety regulation inevitably forces the violation of another.
Autonomous systems routinely face scenarios where strict adherence to every operational rule is mathematically impossible, such as a drone deciding between flying through a dangerously narrow gap or depleting its battery on a longer route.
For a self-driving vehicle, this conflict often materializes when a pedestrian abruptly steps into the road, forcing the car’s navigation system to choose between braking inadequately or swerving across a clearly marked center line into oncoming traffic.
Engineers traditionally optimize these systems using a single mathematical cost function that assigns weighted values to distinct operational goals, including safety, legality, efficiency, and passenger comfort.
Under this model, the robot calculates an aggregated total score for every possible action and selects the highest-scoring maneuver, effectively treating all objectives as interchangeable variables.
If efficiency is weighted too heavily within the algorithm, an autonomous vehicle might execute aggressive maneuvers that compromise passenger security to achieve a faster route.
Tichakorn Wongpiromsarn, associate professor of computer science at Iowa State University, noted that simply adjusting the mathematical weights to encourage cautious behavior does not resolve the underlying architectural flaw.
“In this scenario, safety is being treated as just another factor to trade off,” Wongpiromsarn said. “If safety truly comes first, you can’t capture that with a single weight.”
Wongpiromsarn added that safety should never be balanced against secondary goals, but rather treated as a hard limit that the system refuses to cross.
Alternative robotic designs rely on rigid categories of “hard” and “soft” constraints, dictating that hard constraints must be prioritized regardless of the operational cost.
However, Wongpiromsarn explained that this binary practice fails entirely when a vehicle faces a scenario where a hard constraint—such as preventing physical harm to a pedestrian—cannot be satisfied.
Confronted with a choice between hitting a pedestrian or colliding with another vehicle, a hard-constraint architecture offers no guidance and simply declares the situation unsolvable, paralyzing the system precisely when immediate action is required.
The rulebooks framework, developed alongside Konstantin Slutsky, assistant professor of mathematics at Iowa State, and Emilio Frazzoli, professor of dynamic systems and control at ETH Zürich, resolves this paralysis.
By establishing a strict hierarchy of goals, the system systematically compares unavoidable violations and executes the least harmful action available.
Replacing blended weights with ranked priorities allows autonomous systems to mirror human ethical reasoning under pressure.
“People typically follow the most important rules first and only consider lower-priority goals once the critical ones are met or proven impossible,” Wongpiromsarn said.
Slutsky explained that the rulebooks structure permits a gradual specification of priorities across different regulatory jurisdictions and corporate engineering teams.
Lawmakers can establish non-negotiable base priorities, such as avoiding human harm, while manufacturers retain the flexibility to rank secondary objectives like lane positioning or curb distance.
“Our hope is that this approach supports compliance without over-restricting,” Slutsky said. “Everyone follows the same core rules, but companies still have the freedom to innovate and design their own behavior.”
This hierarchical transparency provides a critical tool for post-incident analysis by regulators, engineers, and courts.
Instead of outputting a single opaque numerical score, the system identifies all optimal actions under a prioritized set of rules, clearly demonstrating whether a machine adhered to its most critical mandates.
“After a crash, near-miss or regulatory review, understanding a machine’s reasoning can be as important as the outcome itself,” Wongpiromsarn said.
In laboratory tests, the researchers demonstrated that rulebook algorithms efficiently generate plans that outperform standard planning methods in complex environments where traditional systems break down.
The study indicates that rulebooks can serve as a universal translator for disparate robot-control methodologies, integrating logical directives with constraint-based approaches.
Directives such as yielding to pedestrians can be processed alongside continuous optimization goals like minimizing travel time, eliminating the friction between competing technical architectures.
As artificial intelligence assumes greater decision-making authority in healthcare, transportation, and public safety, hierarchical rulebooks offer a defensible template for automated judgment calls.
Future iterations of the framework are expected to guide international regulatory compliance standards, ensuring autonomous systems can mathematically justify their actions when a flawless outcome proves impossible.
